WinCLIP Reproduced

Independent reproduction workspace for WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation (CVPR 2023).

This repository is part of a focused visual anomaly-detection reproduction series around CLIP-based industrial anomaly detection. The goal is to reproduce the major methods compared with AF-CLIP, document what matches, and keep transparent debugging records when a metric does not match the original paper exactly.

Current status

The final reported zero-shot results use the Accurate-WinCLIP reference implementation with ViT-B-16-plus-240 and the LAION-400M checkpoint. The reproduced numbers match the Accurate-WinCLIP reference tables.

Dataset	Shot	Pixel AUROC	Pixel AUPRO	Image AUROC	Image AP	Status
MVTec-AD	0	82.3	61.9	90.4	95.6	Reference reproduced
VisA	0	73.2	51.0	75.5	78.7	Reference reproduced

All values are percentages and macro-averaged across categories.

What is included

• Dataset loaders for MVTec-AD and VisA.
• A local prototype WinCLIP implementation.
• An Anomalib-backed diagnostic runner.
• Accurate-WinCLIP reference zero-shot logs for MVTec-AD and VisA.
• Aggregation script for paper/reference comparison.
• Documentation of the Anomalib pixel-localization diagnostic.

Reproduction results

Aggregate table:

results/summary.csv

Raw reference logs:

results/raw/mvtec_accurate_winclip_zs.log
results/raw/visa_accurate_winclip_zs.log

Reference note:

docs/REFERENCE_REPRODUCTION.md

The Anomalib diagnostic note is retained at:

docs/WINCLIP_PIXEL_DIAGNOSTIC.md

Setup

conda env create -f environment.yml
conda activate winclip
pip install anomalib open_clip_torch

The server used for the current reproduction had no direct access to HuggingFace, so the OpenCLIP checkpoint was downloaded separately and provided locally.

Required checkpoint:

timm/vit_base_patch16_plus_clip_240.laion400m_e31/open_clip_pytorch_model.bin

Local SHA256 used:

fa8eec9aff58e9215b9b44a977038179712694d3fc3a73eba62546bcff13deb3

The checkpoint is not redistributed in this repository.

Datasets

Set:

export WINCLIP_DATA_ROOT=/path/to/winclip_data

Expected layout:

WINCLIP_DATA_ROOT/
├── mvtec_ad/
└── visa/

Run

Single Anomalib-backed diagnostic category:

CUDA_VISIBLE_DEVICES=0 python scripts/run_winclip_anomalib.py \
  --dataset mvtec_ad \
  --category bottle \
  --shot 0 \
  --output_dir results_anomalib_zs

Full Anomalib diagnostic sweep:

python scripts/run_all_anomalib.py \
  --datasets mvtec_ad visa \
  --shots 0 \
  --output_dir results_anomalib_zs

Aggregate:

python scripts/aggregate_results.py --results_dir results_anomalib_zs --shots 0

Notes on implementation paths

The earlier Anomalib 2.6.0 path produced near image-level reproduction but unstable pixel-localization behavior under the current OpenCLIP stack. The final reported numbers therefore use the Accurate-WinCLIP reference implementation, while the Anomalib diagnostic files remain in the repository for transparency.

Reproduction series

Focused AF-CLIP comparison-chain reproduction:

• AF-CLIP reproduced
• AnomalyCLIP reproduced
• WinCLIP reproduced — Accurate-WinCLIP reference implementation results
• VAND / APRIL-GAN
• AdaCLIP
• AA-CLIP

References

• Jeong et al., WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation, CVPR 2023.
• Accurate-WinCLIP PyTorch reference implementation.
• Radford et al., Learning Transferable Visual Models From Natural Language Supervision, ICML 2021.
• Bergmann et al., The MVTec Anomaly Detection Dataset, CVPR 2019.
• Zou et al., Spot-the-Difference Self-supervised Pre-training for Anomaly Detection and Segmentation, ECCV 2022.

License

The independently written wrapper code and documentation in this repository are released under the MIT License. This license does not apply to WinCLIP, Accurate-WinCLIP, Anomalib, OpenCLIP, CLIP checkpoints, or the benchmark datasets.